The He initialization method is a variation of the Xavier method that's more suitable for ReLU activation functions because it compensates for the fact that this function returns zero for half of the definition domain. This method of weight calculation relies on a probability distribution with the following variance:

There are also other methods of weight initialization. Which one you choose is usually determined by the problem being solved, the network topology, the activation functions being used, and the loss function. For example, for recursive networks, the orthogonal initialization method can be used. We'll provide a concrete programming example of neural network initialization in Chapter 12, Exporting and Importing Models. 

In the previous sections, we looked at the basic components of artificial neural networks, which are common to almost all types of networks. In the next section, we will discuss the features of convolutional neural networks that are often used for image processing.